Remote sensing image specific camouflage target identification method and system based on unsupervised self-adaption
By adopting an unsupervised adaptive domain adversarial neural network in remote sensing image recognition, the problem of insufficient accuracy and real-timeness of camouflage target recognition in small sample unlabeled optical remote sensing image data is solved, and more efficient cross-domain detection performance and recognition accuracy are achieved.
Patent Information
- Application Number
- CN202411842031.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-04-29
AI Technical Summary
The prior art has shortcomings in the acquisition and accurate identification of high-value target information, especially in small samples without labeling optical remote sensing image data, the accuracy and real-time nature of camouflage target recognition are difficult to meet the needs of confrontation tasks.
A domain adversarial neural network based on unsupervised adaptive is adopted to train the remote sensing image target recognition model through unsupervised adaptive training method, and the data fully labeled from the source domain and the data without labeled from the target domain are used for joint training, reducing the need for data annotation in the target domain.
It improves the cross-domain detection performance of optical remote sensing images, improves the accuracy and effectiveness of camouflage target recognition, reduces the impact on environmental factors, and reduces the need for subjective dependence on discriminators.
Smart Images
Figure CN119904613A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target positioning and recognition, and in particular to a remote sensing image specific camouflage target recognition method and system based on unsupervised self-adaptation. Background Art
[0002] Target camouflage and identification have always been research hotspots in the field of information confrontation. The struggle between the two revolves around the difference between the target and the background, and the difference between the target and the background is the fundamental reason for the target to be exposed. Common adversarial targets use camouflage paint, camouflage nets, and other means to change the surface characteristics of the target by using color, texture, and various patterns to achieve the effect of blending into the surrounding complex environment and achieving the purpose of invisibility. Some camouflage effects are even unrecognizable to the human eye. For example, in certain background environments, camouflage nets can achieve the effect of "different objects with the same spectrum" between the target and the background, which poses a great challenge to typical target identification. Camouflaged object detection (COD) can better solve the above problems. It aims to quickly and accurately detect targets hidden in complex backgrounds.
[0003] Satellite remote sensing is an important means of reconnaissance because of its large-scale synchronous observation, high timeliness, and few restrictions on information acquisition, especially without national boundaries. Remote sensing technology can achieve all-weather, all-day, all-round, and highly dynamic reconnaissance, and can accurately identify and locate typical camouflaged targets. Compared with the rapid development of imaging technology, the recognition of typical camouflaged targets based on high-resolution optical remote sensing images is still in its infancy, and the ability to obtain high-value target information is weak. It is still difficult to meet the application requirements of confrontation tasks in terms of accuracy and real-time performance. To this end, how to use advanced artificial intelligence technology to realize the recognition of camouflaged targets in small sample unlabeled optical remote sensing image data is a key technical problem that needs to be solved urgently. Summary of the invention
[0004] To this end, the present invention provides a method and system for recognizing specific camouflaged targets in remote sensing images based on unsupervised adaptation, which improves the cross-domain detection performance of optical remote sensing images, greatly reduces the demand for target domain data annotation in real environments, and improves the accuracy and effectiveness of camouflaged target recognition in specific mission environments.
[0005] According to the design scheme provided by the present invention, on the one hand, a method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation is provided, comprising:
[0006] Acquire the remote sensing image to be detected;
[0007] Inputting the remote sensing image to be detected into a remote sensing image target recognition model, and using the remote sensing image target recognition model to obtain a camouflaged target category label in the image to be detected;
[0008] Among them, the remote sensing image target recognition model is constructed based on a domain adversarial neural network, which includes a feature extractor and a label classifier connected in sequence, and the feature extractor and the label classifier are obtained by training the remote sensing image target recognition training model using an unsupervised adaptive training method.
[0009] As the remote sensing image specific camouflage target recognition method based on unsupervised adaptation of the present invention, further, the remote sensing image target recognition training model is trained by using an unsupervised adaptive training method, comprising:
[0010] Acquire a source domain training data set and a target domain training data set, wherein the source domain training data set includes source domain training remote sensing images and corresponding camouflaged target category labels, and the target domain training data set includes target domain training remote sensing images;
[0011] Constructing a remote sensing image target recognition training model based on a domain adversarial network, wherein the remote sensing image target recognition training model includes a first feature extractor, a first label predictor, and a first domain classifier, and a gradient flipping layer for back propagation learning is provided between the first feature extractor and the first domain classifier;
[0012] A first feature extractor is used to extract feature information of remote sensing images in a source domain training data set and a target domain training data set, a first label predictor is used to classify features of remote sensing images from the source domain training data set, a first domain classifier is used to classify sources of feature information of remote sensing images, a remote sensing image target recognition training model is trained based on the first label predictor and the first domain classifier training loss function, and a trained first feature extractor and a first label classifier are obtained, so that the trained first feature extractor and the first label classifier are used as feature extractors and label classifiers in the remote sensing image target recognition model.
[0013] As the remote sensing image specific camouflage target recognition method based on unsupervised self-adaptation of the present invention, further, a source domain training data set is obtained, including:
[0014] Collect optical remote sensing image data;
[0015] Marking specific camouflaged targets in collected optical remote sensing image data;
[0016] The target camouflage evaluation system is used to rate and rank the camouflage effects of specific targets, and a source domain training dataset is constructed based on the rating and ranking results.
[0017] As the remote sensing image specific camouflage target recognition method based on unsupervised self-adaptation of the present invention, further, the camouflage effect of the specific target is ranked by using the target camouflage evaluation system, including:
[0018] Setting characteristic indicators of camouflage effect of specific mission targets in optical remote sensing images, and determining weights of each characteristic indicator using a dynamic weight allocation algorithm, wherein the characteristic indicators include image brightness indicator, image color indicator, image texture indicator and target edge contour indicator;
[0019] The similarity of specific targets in optical remote sensing images is obtained by weighted fusion based on the weights of feature indicators, and a multi-indicator target camouflage evaluation system is established based on the similarity.
[0020] According to the multi-index target camouflage evaluation system, the recognition difficulty of specific targets marked in optical remote sensing images is finely graded and sorted to obtain an optical remote sensing image sample set for camouflaged target recognition.
[0021] As the remote sensing image specific camouflage target recognition method based on unsupervised self-adaptation of the present invention, further, the weight of each feature index is determined by using a dynamic weight allocation algorithm, including:
[0022] Formulate a weight update strategy based on the influencing factors of the characteristic indicators on the camouflaged target recognition task, wherein the influencing factors include the influence degree of each characteristic indicator, the contribution of each characteristic indicator and the interaction relationship between each characteristic indicator;
[0023] Based on the weight update strategy and using the dynamic weight allocation algorithm, the weight of each feature indicator is determined.
[0024] As the remote sensing image specific camouflage target recognition method based on unsupervised adaptation of the present invention, further, a remote sensing image target recognition training model is trained based on the first label predictor and the first domain classifier training loss function, comprising:
[0025] The first domain classifier training loss function is set based on binary cross entropy;
[0026] The remote sensing image object recognition training model is trained by minimizing the first label predictor training loss and maximizing the domain classifier training loss.
[0027] As the remote sensing image specific camouflage target recognition method based on unsupervised adaptation of the present invention, further, the first domain classifier training loss function is expressed as: Among them, N S and N T Represent the number of samples from the source domain and the target domain respectively, They represent the i-th sample from the source domain and the j-th sample from the target domain, respectively.
[0028] On the other hand, the present invention also provides a remote sensing image specific camouflage target recognition system based on unsupervised adaptation, comprising: an image acquisition module and an image recognition module, wherein:
[0029] An image acquisition module is used to acquire the remote sensing image to be detected;
[0030] An image recognition module is used to input the remote sensing image to be detected into a remote sensing image target recognition model, and use the remote sensing image target recognition model to obtain a camouflaged target category label in the image to be detected;
[0031] Among them, the remote sensing image target recognition model is constructed based on a domain adversarial neural network, which includes a feature extractor and a label classifier connected in sequence, and the feature extractor and the label classifier are obtained by training the remote sensing image target recognition training model using an unsupervised adaptive training method.
[0032] Beneficial effects of the present invention:
[0033] The present invention uses fully labeled source domain data and unlabeled target domain data to jointly train a model based on an unsupervised domain adaptation method, so that the model can achieve good generalization performance in the target domain without the need for target domain labeling, improve the cross-domain detection performance of the model, greatly reduce the demand for target domain data labeling in a real environment, and use objective standard specific target camouflage effect evaluation indicators to label source domain data, reduce subjective dependence on identification personnel, reduce the impact of environmental factors, improve model training effects, and facilitate deployment and implementation in actual remote sensing image target detection and recognition scenarios, and has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a schematic diagram of a specific camouflaged target recognition process based on unsupervised self-adaptation in a remote sensing image in an embodiment;
[0035] Figure 2 Schematic diagram of the domain adversarial neural network DANN structure in the embodiment. DETAILED DESCRIPTION
[0036] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.
[0037] Remote sensing image target recognition is an important branch of computer vision tasks. It is similar to the method of image target recognition in natural scenes. However, remote sensing images are significantly different from natural images in terms of imaging mode, image quality, image size, etc. The type, perspective, background, etc. of the targets in the images are also significantly different from those in natural images, which makes remote sensing image target recognition more difficult. Due to the difference in types from natural background images, when applying target recognition and detection methods to remote sensing images, certain improvements need to be made to adapt to the characteristics of remote sensing images, so as to achieve better target recognition effects. For an embodiment of the present invention, see Figure 1As shown, a method for recognizing specific camouflaged targets in remote sensing images based on unsupervised adaptation is provided, comprising:
[0038] S101, obtaining a remote sensing image to be detected;
[0039] S102, inputting the remote sensing image to be detected into a remote sensing image target recognition model, and using the remote sensing image target recognition model to obtain a camouflaged target category label in the image to be detected;
[0040] Among them, the remote sensing image target recognition model is constructed based on a domain adversarial neural network, which includes a feature extractor and a label classifier connected in sequence, and the feature extractor and the label classifier are obtained by training the remote sensing image target recognition training model using an unsupervised adaptive training method.
[0041] Adversarial-based domain adaptation methods use adversarial objectives for training. Adversarial learning usually includes a feature extractor and a domain classifier. The domain classifier focuses on distinguishing the source domain from the target domain, while the goal of the feature extractor is to learn domain-invariant features to deceive the domain classifier. In this embodiment, the remote sensing image target recognition model is trained based on the domain adversarial network, which greatly reduces the need for target domain data annotation in the real environment and improves the generalization and robustness of the model.
[0042] Specifically, the unsupervised adaptive training method is used to train the remote sensing image target recognition training model, which can be designed to include:
[0043] Acquire a source domain training data set and a target domain training data set, wherein the source domain training data set includes source domain training remote sensing images and corresponding camouflaged target category labels, and the target domain training data set includes target domain training remote sensing images;
[0044] Constructing a remote sensing image target recognition training model based on a domain adversarial network, wherein the remote sensing image target recognition training model includes a first feature extractor, a first label predictor, and a first domain classifier, and a gradient flipping layer for back propagation learning is provided between the first feature extractor and the first domain classifier;
[0045] A first feature extractor is used to extract feature information of remote sensing images in a source domain training data set and a target domain training data set, a first label predictor is used to classify features of remote sensing images from the source domain training data set, a first domain classifier is used to classify sources of feature information of remote sensing images, a remote sensing image target recognition training model is trained based on the first label predictor and the first domain classifier training loss function, and a trained first feature extractor and a first label classifier are obtained, so that the trained first feature extractor and the first label classifier are used as feature extractors and label classifiers in the remote sensing image target recognition model.
[0046] By collecting satellite remote sensing images containing adversarial scenes, manually annotating specific adversarial targets in the images, and rating and ranking the camouflage effects of the camouflaged targets under the guidance of a camouflage evaluation system, a source domain training dataset is established; and unsupervised domain adaptation technology is used to improve the versatility of the target recognition network.
[0047] Specifically, the source domain training dataset is obtained, which can be designed to include:
[0048] Collect optical remote sensing image data;
[0049] Marking specific camouflaged targets in collected optical remote sensing image data;
[0050] The target camouflage evaluation system is used to rate and rank the camouflage effects of specific targets, and a source domain training dataset is constructed based on the rating and ranking results.
[0051] Among them, using the target camouflage evaluation system to rank the camouflage effects of specific targets may include:
[0052] Setting characteristic indicators of camouflage effect of specific mission targets in optical remote sensing images, and determining weights of each characteristic indicator using a dynamic weight allocation algorithm, wherein the characteristic indicators include image brightness indicator, image color indicator, image texture indicator and target edge contour indicator;
[0053] The similarity of specific targets in optical remote sensing images is obtained by weighted fusion based on the weights of feature indicators, and a multi-indicator target camouflage evaluation system is established based on the similarity.
[0054] According to the multi-index target camouflage evaluation system, the recognition difficulty of specific targets marked in optical remote sensing images is finely graded and sorted to obtain an optical remote sensing image sample set for camouflaged target recognition.
[0055] The source domain data labels are annotated using objective standard specific target camouflage effect evaluation indicators to reduce the reliance on the subjective judgment of personnel, reduce the impact of environmental factors, and improve the model training effect.
[0056] Specifically, a weight update strategy can be formulated based on the influencing factors of the feature indicators on the camouflaged target recognition task, wherein the influencing factors include the influence degree of each feature indicator, the contribution of each feature indicator and the interaction relationship between each feature indicator; based on the weight update strategy and using the dynamic weight allocation algorithm, the weight of each feature indicator is determined.
[0057] In the scope of optical camouflage, the factors affecting the camouflage effect include target brightness coefficient, color, texture, geometric characteristics and spectral reflection characteristics. Among them, the spectral reflection characteristics are related to the reflection of the color characteristics of the object surface. The spectral reflection characteristics of the target can be reflected by extracting the color characteristics, and the geometric characteristics can be reflected by extracting the edge contour characteristics of the target itself. In terms of the human eye visual perception system, the evaluation of the camouflage effect under visible light mainly affects the subjective perception of human vision. The target recognition characteristics under visible light include brightness, color, edge, shape, boundary contrast and texture difference. The selection of index features should be based on the comprehensiveness of the overall index, while ensuring the independence of individual indicators. According to current research, the brightness index, color index and texture index have greater visual stimulation to the human eye. The edge of the object can reflect its external shape characteristics. At the same time, whether the object can be directly integrated into the background at the boundary is the decisive factor in whether the target can be discovered. Therefore, these four independent indicators can be selected to construct an evaluation index system.
[0058] Generally speaking, the brightness of the target will directly affect the fusion effect between the target and the background. The smaller the difference between the overall brightness of the target and the surrounding background environment, the better the target can be integrated into the environment, achieving a good camouflage purpose. In the extraction of the target brightness index, the first thing to do is to choose a suitable color space. The color of the normally acquired color image is in the RGB color space, which is not suitable for calculating and processing the image. Combined with specific feature extraction, the CIE Lab color space is often used in the processing of brightness features. The Lab color space can fully retain the brightness information. The CIE Lab mode is a color mode based on the physiological characteristics of the human eye, which can specifically express the visual perception of the human eye. It contains three basic elements: L (brightness), a (green and red color channel), and b (blue and yellow color channel). Brightness is a separate information channel, so it can be processed separately to extract the brightness index of the image.
[0059] The relationship between the brightness value L and the XYZ color space conversion is as follows:
[0060]
[0061] Wherein, the reference white light radiation brightness value is Y0=100.
[0062] As a widely used image feature, color feature extraction is relatively easy compared to other features, and is less affected by geometric transformations and has strong robustness. As one of the widely used color feature extraction methods, the color histogram is a method that describes the proportion of different colors in an image to the color of the entire image through statistical distribution characteristics. This method can only describe the proportion of the number of pixels of the same color, and cannot record the spatial position of the color, so it is not affected by the geometric transformation of the image. Normalizing the histogram can facilitate the statistical difference in color features, which refers to the ratio of the number of pixels of the same color to the total number of pixels, as shown in the following formula:
[0063] H(k)=n k / N,k=0,1,…L-1
[0064] Where k is the color value; L is the number of all color values; n k is the number of pixels with color value k; N is the total number of pixels.
[0065] Texture in an image is often used to describe the grayscale changes of each pixel in spatial distribution, including visual features such as roughness, periodicity, and directionality. Whether the target and background can be well integrated and whether the overall distribution characteristics are coordinated visually can be described by comparing the texture differences between the two. When the texture difference between the target and the background is within a certain range, it is considered that the texture similarity between the target and the background is high. The gray level co-occurrence matrix can well describe the spatial distribution of texture features and calculate the texture similarity between images.
[0066] Gray-level co-occurrence matrix is a second-order statistical method that can not only reflect the spatial distribution difference of pixel gray levels, but also describe the spatial information of the relative position between pixels. Pixels in an image can be expressed by position coordinates and gray levels. Suppose the image has X pixels in the horizontal direction, Y pixels in the vertical direction, and the gray level is w. The image information expression is as follows:
[0067]
[0068] The extraction of the boundary line between the target and the background is called edge detection. The commonly used edge operators are the classic template operator method and the linear filtering method. The classic template operator belongs to a nonlinear filtering algorithm with simple calculations and is suitable for edge extraction of images with simple shapes. Linear filtering methods include the Canny operator, wavelet edge detection operator, LOG operator, etc. Among them, the Canny operator is more commonly used in edge detection, and has a better effect on edge contour extraction, which is suitable for edge extraction of images with complex noise. In the embodiment of this case, the Canny operator can be used to obtain the contour edge of the target in the remote sensing image, the image can be smoothed by using a Gaussian filter to filter out the noise in the remote sensing image, and the Sobel operator is used to calculate the horizontal and vertical gradient values of each pixel in the image, and the gradient amplitude and direction of each pixel are obtained according to the gradient value; non-maximum suppression is performed on the obtained gradient amplitude image to narrow the edge width, making the edge more refined and clear; according to the set high threshold and low threshold, the pixels in the gradient amplitude image are divided into three categories: strong edge, weak edge and non-edge, and the complete edge image is obtained by connecting the strong edge pixels and the weak edge pixels adjacent to them.
[0069] The remote sensing image target recognition training model is trained based on the first label predictor and the first domain classifier training loss function, and can be designed to include:
[0070] The first domain classifier training loss function is set based on binary cross entropy;
[0071] The remote sensing image object recognition training model is trained by minimizing the first label predictor training loss and maximizing the domain classifier training loss.
[0072] Domain-adversarial Neural Networks (DANN) implements domain adaptation based on adversarial learning in an adversarial manner. Figure 2 As shown in the figure, it mainly includes feature extractors, domain classifiers, and classifiers related to real tasks. The feature extractor is a network shared by the two branches. The gradient reversal layer enables the former to try its best to learn domain-invariant features so that the domain classifier cannot distinguish whether the features come from the source domain or the target domain, thus achieving the alignment of the spatial distributions of the two domains. In the DANN network, the gradient reversal layer (GRL) is integrated into the network structure to calculate the minimax loss. The function of the gradient reversal layer is to multiply the gradient by a sign during back propagation to ensure that the feature extractor will try its best to learn domain-invariant features to deceive the discrimination results of the domain classifier. Among them, the domain classifier uses binary cross entropy as the loss function, which is specifically expressed as follows:
[0073]
[0074] Furthermore, based on the above method, an embodiment of the present invention also provides a remote sensing image specific camouflage target recognition system based on unsupervised adaptation, comprising: an image acquisition module and an image recognition module, wherein:
[0075] An image acquisition module is used to acquire the remote sensing image to be detected;
[0076] An image recognition module is used to input the remote sensing image to be detected into a remote sensing image target recognition model, and use the remote sensing image target recognition model to obtain a camouflaged target category label in the image to be detected;
[0077] Among them, the remote sensing image target recognition model is constructed based on a domain adversarial neural network, which includes a feature extractor and a label classifier connected in sequence, and the feature extractor and the label classifier are obtained by training the remote sensing image target recognition training model using an unsupervised adaptive training method.
[0078] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0079] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0080] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.
[0081] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.
[0082] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. A method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation, characterized in that: Include: Acquire the remote sensing image to be detected; Inputting the remote sensing image to be detected into a remote sensing image target recognition model, and using the remote sensing image target recognition model to obtain a camouflaged target category label in the image to be detected; Among them, the remote sensing image target recognition model is constructed based on a domain adversarial neural network, which includes a feature extractor and a label classifier connected in sequence, and the feature extractor and the label classifier are obtained by training the remote sensing image target recognition training model using an unsupervised adaptive training method.
2. The method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation according to claim 1, characterized in that: The remote sensing image target recognition training model is trained using an unsupervised adaptive training method, including: Acquire a source domain training data set and a target domain training data set, wherein the source domain training data set includes source domain training remote sensing images and corresponding camouflaged target category labels, and the target domain training data set includes target domain training remote sensing images; Constructing a remote sensing image target recognition training model based on a domain adversarial network, wherein the remote sensing image target recognition training model includes a first feature extractor, a first label predictor, and a first domain classifier, and a gradient flipping layer for back propagation learning is provided between the first feature extractor and the first domain classifier; A first feature extractor is used to extract feature information of remote sensing images in a source domain training data set and a target domain training data set, a first label predictor is used to classify features of remote sensing images from the source domain training data set, a first domain classifier is used to classify sources of feature information of remote sensing images, a remote sensing image target recognition training model is trained based on the first label predictor and the first domain classifier training loss function, and a trained first feature extractor and a first label classifier are obtained, so that the trained first feature extractor and the first label classifier are used as feature extractors and label classifiers in the remote sensing image target recognition model.
3. The method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation according to claim 2, characterized in that: Get the source domain training dataset, including: Collect optical remote sensing image data; Marking specific camouflaged targets in collected optical remote sensing image data; The target camouflage evaluation system is used to rate and rank the camouflage effects of specific targets, and a source domain training dataset is constructed based on the rating and ranking results.
4. The method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation according to claim 3 is characterized in that: The target camouflage evaluation system is used to rank the camouflage effects of specific targets, including: Setting characteristic indicators of camouflage effect of specific mission targets in optical remote sensing images, and determining weights of each characteristic indicator using a dynamic weight allocation algorithm, wherein the characteristic indicators include image brightness indicator, image color indicator, image texture indicator and target edge contour indicator; The similarity of specific targets in optical remote sensing images is obtained by weighted fusion based on the weights of feature indicators, and a multi-indicator target camouflage evaluation system is established based on the similarity. According to the multi-index target camouflage evaluation system, the recognition difficulty of specific targets marked in optical remote sensing images is finely graded and sorted to obtain an optical remote sensing image sample set for camouflaged target recognition.
5. The method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation according to claim 4, characterized in that: The dynamic weight allocation algorithm is used to determine the weight of each feature indicator, including: Formulate a weight update strategy based on the influencing factors of the characteristic indicators on the camouflaged target recognition task, wherein the influencing factors include the influence degree of each characteristic indicator, the contribution of each characteristic indicator and the interaction relationship between each characteristic indicator; Based on the weight update strategy and using the dynamic weight allocation algorithm, the weight of each feature indicator is determined.
6. The method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation according to claim 2, characterized in that: Training a remote sensing image target recognition training model based on a first label predictor and a first domain classifier training loss function includes: setting a first domain classifier training loss function according to binary cross entropy; The remote sensing image object recognition training model is trained by minimizing the first label predictor training loss and maximizing the domain classifier training loss.
7. The method for identifying specific camouflaged targets in remote sensing images based on unsupervised adaptation according to claim 2 or 6, characterized in that: The first domain classifier training loss function is expressed as: Among them, N S and N T Represent the number of samples from the source domain and the target domain respectively, and They represent the i-th sample from the source domain and the j-th sample from the target domain, respectively.
8. A remote sensing image specific camouflage target recognition system based on unsupervised adaptation, characterized in that: It includes: image acquisition module and image recognition module, wherein: An image acquisition module is used to acquire the remote sensing image to be detected; An image recognition module is used to input the remote sensing image to be detected into a remote sensing image target recognition model, and use the remote sensing image target recognition model to obtain a camouflaged target category label in the image to be detected; Among them, the remote sensing image target recognition model is constructed based on a domain adversarial neural network, which includes a feature extractor and a label classifier connected in sequence, and the feature extractor and the label classifier are obtained by training the remote sensing image target recognition training model using an unsupervised adaptive training method.
9. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 can be implemented.